Width: 390 × 844 (also applies at desktop).

Start a practice run, bail out, select Hash rhythm with sound off and click Start rhythm.

After 500 ms click Enable game sound and wait 2.4 seconds.

Actual: the toggle has aria-pressed="true" and is named Mute game sound, but instrumentation around AudioContext and oscillator.start records 0 contexts and 0 tones.

Stop and restart without changing the toggle; after 2.4 seconds the same instrumentation records 1 context and 2 tones.

Expected: enabling sound during the first session produces its subsequent beat tones without restarting.

Width: 390 × 844 (the rule applies through 680px).

Open the page, select Hash rhythm or Round predictor, then press the header Stream view control.

Actual measured state: .scene has display:none; canvas.isVisible() is false; .game-controls is also hidden.

The screenshot shows only the habitat heading and telemetry above the Copy stream URL control, with no wheel.

Exit stream view, select Wheel run and re-enter Stream view: .scene is display:block and canvas.isVisible() is true.

Expected: the canvas stays visible in stream view from either tab; API availability does not affect this reproduction.

Width: 390 × 844.

Use controlled API responses accepted by the current adapters: status {rpm:0,running:false,block_height:1140}, a valid economy response, and settlements containing one height/work_units/hash entry.

Select Hash rhythm and click Start rhythm.

After 300 ms make the next settlements request return HTTP 503 and press Retry live telemetry (an automatic 15-second refresh reproduces the same state).

Actual: the button still reads "Tap the pulse" but isDisabled() is true; instructions remain "0 of 16 pulses hit · Tap or press Space"; Stop rhythm remains visible and the beat timer continues.

Expected: the active round remains tappable using its stored hash, while only starting another round requires available history.

These are explicitly controlled failure inputs, not a claim that the public API returned 503.

Width: 390 × 844; independently repeated at 1440 × 844.

Serve the production export.

Intercept the original API URLs with HTTP-200 JSON: /v1/status {"rpm":0,"running":false,"block_height":1140}; /v1/economy/state {"total_work_units":50000,"total_settlements":101}; /v1/settlements?range=24h {"settlements":[{"height":1140,"work_units":200,"hash":"0x0000000000000000000000000000000000000000000000000000000000000000"}]}.

Instrument AudioContext.createOscillator().start to record performance.now().

Select Hash rhythm, click Start rhythm with sound off, wait 5,000 ms, click Enable game sound, and observe for 250 ms.

This hash schedules beats 600 ms apart (1,600, 2,200, 2,800 ms, etc.).

Actual: seven oscillator starts arrive in that observation window; consecutive gaps are 23, 25, 25, 25, 25 and 25 ms at both widths, replaying past beats.

The same catch-up also occurred with normal public settlement hashes when muting then re-enabling sound.

Expected: discard missed silent beats and play only the next scheduled tone in sync with the existing visual/tap timeline, with no 25 ms catch-up burst.

These API responses are controlled test inputs, not a claim that the public service returns an all-zero hash.